CN105095833B - Network construction method, recognition method and system for face recognition - Google Patents
Network construction method, recognition method and system for face recognition Download PDFInfo
- Publication number
- CN105095833B CN105095833B CN201410193260.5A CN201410193260A CN105095833B CN 105095833 B CN105095833 B CN 105095833B CN 201410193260 A CN201410193260 A CN 201410193260A CN 105095833 B CN105095833 B CN 105095833B
- Authority
- CN
- China
- Prior art keywords
- matrix
- value
- layer
- picture
- face
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 78
- 238000010276 construction Methods 0.000 title claims description 7
- 239000011159 matrix material Substances 0.000 claims abstract description 120
- 238000012549 training Methods 0.000 claims abstract description 67
- 238000013528 artificial neural network Methods 0.000 claims abstract description 38
- 238000000605 extraction Methods 0.000 claims abstract description 25
- 239000000284 extract Substances 0.000 claims abstract description 23
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 22
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 16
- 239000000203 mixture Substances 0.000 claims abstract description 5
- 210000002569 neuron Anatomy 0.000 claims description 31
- 230000006870 function Effects 0.000 claims description 22
- 230000008569 process Effects 0.000 claims description 22
- 210000004205 output neuron Anatomy 0.000 claims description 15
- 238000012545 processing Methods 0.000 claims description 12
- 230000013016 learning Effects 0.000 claims description 11
- 230000035945 sensitivity Effects 0.000 claims description 8
- 230000001537 neural effect Effects 0.000 claims description 7
- 238000009795 derivation Methods 0.000 claims description 3
- 235000013399 edible fruits Nutrition 0.000 claims 1
- 238000005070 sampling Methods 0.000 description 25
- 238000004458 analytical method Methods 0.000 description 15
- 238000013135 deep learning Methods 0.000 description 15
- 238000005516 engineering process Methods 0.000 description 6
- 238000013461 design Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 5
- 230000004913 activation Effects 0.000 description 3
- 230000004927 fusion Effects 0.000 description 3
- 210000004556 brain Anatomy 0.000 description 2
- 238000007796 conventional method Methods 0.000 description 2
- 230000007423 decrease Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000011478 gradient descent method Methods 0.000 description 2
- 230000005764 inhibitory process Effects 0.000 description 2
- 238000003475 lamination Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 210000005036 nerve Anatomy 0.000 description 2
- 230000003252 repetitive effect Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000001360 synchronised effect Effects 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000012806 monitoring device Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention proposes a kind of deep layer network establishing methods estimated for gender classification or age, and the method includes: all training pictures are divided into several groups by step 101)；Step 102) extracts the high-level characteristic of one group of trained picture based on convolutional neural networks, and then obtains the first matrix being made of high-level characteristic vector；The low layer and global characteristics of same group of trained picture are extracted based on artificial neural network simultaneously, and then obtain the second matrix of low-level feature vector composition；The judgement formula of the first matrix, the second matrix and definition based on extraction obtains the result of one group of gender identification or age estimation；The first weight matrix w in judgement formula defined in it1, the second weight matrix w2, bias matrix b and adjust the value of weight beta and updated using error backpropagation algorithm, obtain the final value of these parameters, completion network struction.The judgement formula that the value of the parameter determined when being completed based on network struction is determined carries out the age of face to be identified and the judgement of gender.
Description
Technical field
The present invention relates to computer vision and depth learning technology field, in particular to a kind of network for recognition of face
Construction method, recognition methods and system.
Background technique
Face identifies for computer as one and analyzes all very difficult object, since the 1990s
Just cause the extensive concern of researchers.And the effective human face analysis of success is in intelligent monitoring, video index and population letter
There is huge application prospects again in the fields such as breath statistics.Human face analysis mainly includes the age of the gender identification and face of face
The mean absolute error of estimation, the accuracy rate of Gender Classification and age estimation is the critical index of the two respectively.
Presently, there are human face analysis field correlative study, be all based on artificial " craft " and determine that Feature Descriptor carrys out table
It lets others have a look at face, combining classification device algorithm or regression algorithm expansion.Manually selected feature often expends a large amount of early-stage preparations time,
It with subjectivity, and selects the feature come and is often showed well in certain a kind of data, and expand to other data when property
It can be decreased significantly, generalization ability is weaker.And when practical application, the weak generalization ability of conventional method is just lower in quality
Performance is bad on picture (intense light irradiation picture, there are background interference pictures, wry face side face picture), can not meet the need of practical application
It asks.The bottleneck that research in human face analysis field is limited at present and develop slowly.On the other hand, in recent years, it is based on depth
The method for practising thought achieves great development, provides the support on algorithm to intelligent direction development for computer science.
The basic thought of deep learning is exactly the artificial neural network for constructing deep layer, simulates the study mechanism of human brain, using non-supervisory
The feature of mode " automatic " learning objective object of habit, the feature learnt have hierarchical structure, general from detail to being abstracted
It reads, such feature has more essential portray to data itself.Application of the method for deep learning in many fields all obtains
Breakthrough success, the comprehensive item of image classification speech recognition of the Handwritten Digit Recognition Systems of U.S.'s more banks, Google
Mesh Google Brain, Microsoft full-automatic simultaneous interpretation system be all based on deep learning method realize.Depth at present
The method of study mainly includes that deep layer is sparse from encryption algorithm, deepness belief network algorithm and convolutional neural networks algorithm.Its
Middle convolutional neural networks algorithm all achieves level advanced in the world, such as Face datection in the multiple fields of image procossing, row
People's detection and recognition of face etc.；And deepness belief network algorithm, mainly in field of speech recognition using more, effect is preferable；
The problem of sparse algorithm of coding certainly of deep layer is more applied to Data Dimensionality Reduction class.
Convolutional neural networks essence is a kind of neural network structure of deep layer, and the basic principle and feature of algorithm are networks
Structure is made of two kinds of part and parcels: under the convolutional layer and down-sampling processing unit that convolution processing unit is constituted are constituted
Sample level.Convolutional layer and down-sampling layer form two-dimensional structure by neuron, to be equally two-dimensional structure input picture respectively into
Row process of convolution and down-sampling processing, then repeatedly convolution sum down-sampling, " ideal " until extracting image (are answered according to research
With it needs to be determined that) after feature, then classified or returned or detected.When input picture is N × N size, connect first
Convolutional layer C1, C1In each neuron only receive domain (also referred to as convolution kernel, filter) with one piece of upper one layer part and be connected.
Assuming that the size of convolution kernel is m × m, then C1Layer will input all possible (N-m of picture with the convolution nuclear convolution of this m × m
+ 1) × (N-m+1) pixel of position generates the local feature figure of (N-m+1) × (N-m+1).Input each piece of m of picture
All pixels point and C in × m size area1A neuron is connected in layer, and identical weight is taken in this m × m connection
(i.e. weight shared mechanism).When the connection type using multiple and different weights generates multiple local feature figures, C1Layer just extracts
The different local features of original image out.Then C1The local feature figure of layer is connected to next layer of down-sampling layer S1.Assuming that C1Layer has F1
Characteristic pattern is opened, then corresponding C1Layer also has F1Open down-sampling figure, and and C1The characteristic pattern of layer corresponds.S1In each figure of layer
Each neuron connect one layer on this of one piece of local field, and be not overlapped, then calculate all values in this block region
Value of the average value as sample level.Assuming that C1The size of every characteristic pattern is (N-m+1) × (N-m+1), local bonding pad in layer
Domain size is n × n, then S1The size of each down-sampling figure of layer is (N-m+1)/n × (N-m+1)/n.It realizes in this way
To the down-sampling of upper one layer of characteristic pattern, spatial resolution is reduced.Then S1Layer again with a new convolutional layer C2Layer connection, C2
Layer and S2Layer connection, so intersects repeatedly, determine according to actual needs the number of the number of convolutional layer and down-sampling layer.Finally net
The output of network is referred to as output layer, and the difference according to problem may be the different classification of output, it is also possible to output regression estimation
Probability.
However, although the convolution of convolutional neural networks+down-sampling processing mode can extract the high-rise, abstract of target object
Feature, for achieving good effect when certain classification problems, but have ignored some low layers to the effective feature of classifying
And information.And convolution+down-sampling handles the mainly local feature of object and part relevance for the characteristic pattern description extracted,
Lack the performance to object global feature.When solving gender classification and face age estimation problem, exactly need to face
Comprehensive, multi-level feature extraction and description are carried out, accurate, comprehensive information can be obtained.
Based on above-mentioned, in the method that human face analysis field applies traditional " craft " to determine Feature Descriptor, or it is direct
Using convolutional neural networks (method including existing other deep learnings), their performances and final effect are not able to satisfy
The needs of practical application.The completely new method based on multiple features deep learning of present invention application carries out human face analysis, building training
The network of deep layer, " automatic " learn, extract the feature by different level, comprehensive of face, formed a kind of multiple features (it is high-rise,
Low layer is abstracted, is specific) structure description indicate face.Such multiple features structure is applied to final Gender Classification and age
Extraordinary effect is achieved when estimation.
Still lack such method or system in the prior art.
Summary of the invention
It is an object of the present invention to overcome the performance of the conventional method in the technology of existing human face analysis to can not meet reality
The defect that border application needs, to propose a kind of network establishing method for recognition of face and the people based on the building network
Face recognition method and system.
To achieve the goals above, the present invention provides a kind of deep layer network structure estimated for gender classification or age
Construction method, the method include:
All training pictures are divided into several groups by step 101)；
Step 102) extracts the high-level characteristic of one group of trained picture based on convolutional neural networks, and then obtains by high-level characteristic
First matrix of vector composition；The low layer and global characteristics for extracting same group of trained picture based on artificial neural network simultaneously, into
And obtain the second matrix of low-level feature vector composition；
The first matrix, the second matrix and following judgement formula based on extraction obtain one group of gender identification or age estimation
Result:
O=sigm (w1*hfo+β×w2*lfo+b)
Wherein, hfo indicates the first matrix；Lfo indicates the second matrix；For first group of trained picture above-mentioned formula of input
In the first weight matrix w1, the second weight matrix w2, bias matrix b and adjust weight beta initial value use random initializtion mode
It obtains；For the w in the training picture above-mentioned formula of remaining each group of input1、w2, b and β acquisition methods are as follows: using error it is anti-
Error function J (W, the b of the physical tags matrix Y of court verdict o and each group training picture are calculated to propagation algorithm；β), then pass through
Calculate w1、w2, b and β be to error function J (W, b；Gradient and then undated parameter w β)1、w2, b and β value；
Step 103) inputs one group of trained picture again, and repeats above-mentioned steps to the training picture inputted again
102), until the processing that all groupings are carried out step 102), completes primary training iteration；
All training pictures are reclassified as several groups by step 104), and are repeated to each group repartitioned
Step 102) and step 103) are stated, iteration again is completed；
By several grouping and iterative processings again, until being obtained most when the judgement o of final output meets the condition of setting
Whole parameter w1、w2, b and β value, complete network struction.
Optionally, it carries out further including when low-level feature abstract:
Step 102-11) each Zhang Xunlian picture of one group of input trained picture is converted by two-dimensional graph structure form
For the form of vector, then vector is normalized, obtains the original feature vector of each Zhang Xunlian picture；
Step 102-12) by the original vector input artificial neural network of obtained each Zhang Xunlian picture, and then obtain
One group of reconstruction features vector to get arrive second matrix；Wherein, the artificial neural network includes L layer, and layer and layer it
Between use full connection type, each layer of each neuron is using the activation of sigmoid function.
Optionally, the gender or age identification process for inputting training picture for one specifically include:
Step 102-21) when extraction a trained picture high-level characteristic vector be HfThe high-level characteristic vector of dimension, and
Low-level feature vector is LfWhen the feature vector of dimension, construction one includes " Hf+Lf" joint of a neuron decides by vote layer；
Step 102-22) when for gender identification when, by construction joint voting layer each neuron respectively with output
Two output neurons of layer are connected, and each output neuron is based on the judgement formula and carries out Sex Discrimination, output training
Picture is the probability of sex；When for age estimation, each neuron of joint voting layer and the S of output layer are a defeated
Out neuron be connected, wherein each output neuron correspond to it is one-year-old.
Optionally, the first weight matrix w is updated using the error backpropagation algorithm of following formula1Value:
Wherein, (w1)newIndicate the updated first weight matrix w in error back propagation each time1Value,
(w1)oldThe first weight matrix w before corresponding update1Value, Od indicate output layer sensitivity matrix, the output layer sensitivity matrix
Using error function J (W, b；β) court verdict o is combined to find out；α indicates the learning rate of network, wherein the value of α is initialized as
Then one biggish value is gradually reduced with the increase of training the number of iterations；
The second weight matrix w is updated by following formula2Value:
Wherein, (w2)newIndicate the updated second weight matrix w in error back propagation each time2Value,
(w2)oldThe second weight matrix w before corresponding update2Value.
Optionally, the value of the β is training update method in iteration each time are as follows:
Wherein, βnewIndicate the updated value for adjusting weight beta, β in error back propagation each timeoldIt is corresponding to update
The value of preceding adjusting weight beta；
The part of local derviation is asked in above-mentioned formula to be obtained by following formula:
Wherein, f ' (o) indicates that, to court verdict o derivation, " mean (B (:)) " indicates to be averaged all elements in matrix B
It is worth operation；Matrix B indicates the matrix being made of the value for adjusting weight beta updated in error back propagation each time, should
The ranks number of matrix B and the ranks number of court verdict o are identical.
Deep layer network based on above-mentioned building, the present invention also provides a kind of age for face or the identification sides of gender
Method, the method include:
For extracting the high-level characteristic of face picture to be identified based on convolutional neural networks；
For extracting the low layer and global characteristics of face picture to be identified based on artificial neural network；
For the low-level feature of extraction and high-level characteristic to be inputted to following judgement formula, carries out gender or the age is sentenced
Certainly, court verdict is exported:
O=sigm (w1*hfo+β×w2*lfo+b)
Wherein, the first weight matrix w in above-mentioned formula1, the second weight matrix w2, β and b be deep layer network establishing method
Determining value, hfo indicate that the high-level characteristic vector of the face to be identified extracted, lfo indicate the low layer of the face to be identified extracted
Feature vector, o indicate gender or the court verdict at age.
Low-level feature is extracted using following steps:
After face picture to be identified for that will input carries out flaky process and normalizes, the original spy of face is obtained
Levy vector；
For original feature vector to be inputted artificial neural network, weight is carried out to input vector by multilayer neuronal structure
It builds, obtains a LfThe feature vector of dimension is as the low-level feature vector extracted.
In addition, the system includes the present invention provides a kind of age for face or the identifying system of gender:
High-level characteristic extraction module, for extracting the high-level characteristic of face picture to be identified based on convolutional neural networks；
Low-level feature abstract module, the low layer and the overall situation for extracting face picture to be identified based on artificial neural network are special
Sign；
Judging module neural network based, it is public for the low-level feature of extraction and high-level characteristic to be inputted following judgement
Formula carries out gender or age judgement, exports court verdict:
O=sigm (w1*hfo+β×w2*lfo+b)
Wherein, the first weight matrix w in above-mentioned formula1, the second weight matrix w2, β and b be to complete for training picture
Iteration several times after obtain (the final value of each parameter obtained when i.e. above-mentioned network struction is completed), what hfo indicated to extract
The high-level characteristic vector of face to be identified, lfo indicate that the low-level feature vector of the face to be identified extracted, o indicate gender or year
The court verdict in age.
Optionally, above-mentioned low-level feature abstract module further includes:
Flaky process module, after the face picture to be identified for that will input carries out flaky process and normalizes,
Obtain the original feature vector of face；
Reconstruction features vector obtains module, for original feature vector to be inputted artificial neural network, passes through multilayer nerve
Meta structure rebuilds input vector, obtains a LfThe feature vector of dimension is as the low-level feature vector extracted.
Optionally, above-mentioned judging module further includes:
Joint voting layer module exports a kind of multiple features structure for merging the high-level characteristic and low-level feature that extract；
Output layer module, for carrying out gender judgement or age judgement, each mind using several output neurons
It is made decisions through member based on the judgement formula.
Compared with traditional " craft " determines the method for Feature Descriptor, the technical advantages of the present invention are that:
Deep layer network provided by the invention is capable of the feature of " automatic " study face, and the feature learnt has layering
Structure, such feature to data itself have it is more essential portray, so finally being classified and being returned using such feature
Return effect when estimation also more preferable.Further, compared with the method for existing deep learning, deep layer network of the invention can not only
Learn the higher level of abstraction feature to face, while learning the low layer global characteristics of face, it is comprehensive, multi-level in conjunction with two kinds
Description indicates face, and when classification final in this way and regression estimates, performance of the invention is than being only extracted individual high-level characteristic
Existing deep learning method it is more preferable.The method of this deep learning based on multiple features of the present invention is applied to face point
When analysis field, not only there is extremely strong learning ability, there are also extremely strong generalization abilities.In the high quality mark of test common data sets
When the practical face picture of quasi- face picture, the lower network face picture of quality and monitoring device acquisition, all achieve super
Cross the performance of the prior art.Therefore the present invention can satisfy the needs of practical human face analysis application.
Detailed description of the invention
Fig. 1 is gender classification flow chart provided by the invention；
Fig. 2 is to estimate flow chart at the face age provided by the invention；
Fig. 3 is the structural schematic diagram for the human face analysis system that embodiment of the present invention provides.
Specific embodiment
Now in conjunction with attached drawing, the invention will be further described.
The present invention gives a kind of human face analysis methods based on multiple features deep learning, comprising:
Step 1) carries out Face datection and pretreatment to picture.
Step 2) carries out human face analysis to the face picture that step 1) obtains, and is separately input to gender classification deep layer
Network and face age estimate deep layer network.
The gender of step 3), the face picture of gender classification deep layer network output prediction, male or female；The face age is estimated
How old count the age integer value of the face picture of deep layer network output estimation.
One, the building of gender classification network:
In above-mentioned technical proposal, the step 2) provides a kind of face gender identification method, as shown in Figure 1, i.e. one
Kind is used for the deep layer network system of gender classification, comprising:
Step 2-1) using face picture obtained in step 1) as the input of network, the output of network is people in picture
Gender prediction's value (male or female).Network is mainly made of 4 partial function modules, including high-level characteristic extraction module, low-level feature
Extraction module, fusion feature cascading judgement output module and parameter training module.
In above scheme, the step 2-1) specifically comprise the following steps:
Step 2-1-1), high-level characteristic extract: directly adopt the volume of the convolutional neural networks in existing deep learning method
Lamination and down-sampling layer structure, which are realized, extracts the high-level characteristic for inputting training picture.In the specific implementation, using 3 layers of convolutional layer
C1, C2And C3And two layers of down-sampling layer S1, S2Combination, full connection between layers.
Step 2-1-2), low-level feature abstract: it is synchronous with high-level characteristic extraction module to input training image handle.
Firstly, the form for converting vector by two-dimensional graph structure form for the training picture of the face of input is (referred to as flat
Graduation processing, flat operation).Obtained vector is normalized again after carrying out flat operation to input picture, obtains face
Original feature vector.
Then, original feature vector is connected to and rebuilds network to obtain reconstruction features vector.Network is rebuild to be based on manually
Neural networks principles are built, L layers total.Each layer of neuron output is considered as feature vector, next layer of neuron pair
The vector of upper one layer of output is recompiled, and is indicated and is exported again again after describing, takes full connection between layers
Mode, each neuron activated using sigmoid function.Original feature vector is first connected to the H for rebuilding network1
Layer, H1Layer contains h1A neural unit, it is assumed that input picture size is N × N, then face original feature vector is N × N-dimensional, warp
Cross H1Become h after layer1Dimensional feature vector.It is then followed by and is connected to H2Layer, H2Layer contains h2A neural unit, then feature vector into
One step becomes h2Dimension.And so on, according to actual needs, it is finally coupled to HnLayer, obtains a hnThe feature vector of dimension.Layer with
Connection between layer is expressed with mathematical formulae are as follows:
al+1=sigm (Wl·al+bl) (1)
Wherein " sigm () " indicates that the matrix form of sigmoid function (carries out sigmoid to each element in matrix
Function activation), al+1And alRespectively indicate the matrix form (referred to herein as feature vector) of (l+1) layer and l layers of output, Wl
For connection (l+1) weight matrix of neuron, b between layer and l layerslIndicate l layers of bias matrix.
This hnThe feature vector of dimension is to rebuild output namely the reconstruction features vector of network.Reconstruction features vector picks
In addition to redundancy invalid to Gender Classification in face original feature vector, reduce partial noise interference, it can be preferable
Portray the low layer and global characteristics of face.
Step 2-1-3), cascading judgement output: the high-level characteristic that combined extracting arrives is carried out together with low-level feature finally
Gender judgement, output gender prediction's value (male or female).
The high-level characteristic figure that high-level characteristic extraction unit is got is converted into vector form after carrying out flat operation, with low layer spy
Sign, which extracts obtained reconstruction features vector and links together to constitute, combines voting layer, and joint voting layer still is based on artificial neural network
Network principle is connected with upper one layer.Assuming that high-level characteristic extracts the characteristic pattern that finally obtained G q × q sizes, then be converted into
A G × q × q can be obtained after amount and tie up high-level characteristic vector；Reconstruction features vector is hnDimension；So contain in joint voting layer
(G×q×q+hn) a neuron, high-level characteristic vector sum reconstruction features vector is merged, a kind of multiple features structure is formed
The vector of form.The output of joint voting layer is the (G × q × q+h for the multiple features structure that our whole networks are extractedn) dimension
Feature vector.
Joint voting layer is connected to final output neuron entirely again, and there are two (two class of men and women), outputs for output neuron
Be final result one kind Probability pi, output neuron is activated using sigmoid function, then the probability of every one kind can table
It is shown as:
WhereinIndicate the output of joint voting k-th of neuron of layer,Indicate joint voting i-th of neuron of layer with
The connection weight of k-th of neuron of output layer,It is biased for output layer is corresponding.
Because of multiple training picture (parameter training part has respective description) of the input every time of whole network, thus it is of the invention
The matrix form of the judgement output result of the network of definition are as follows:
O=sigm (w1*hfo+w2*lfo+b) (3)
Wherein, o indicates the court verdict (court verdict that each column indicate a sample) of network output；w1Indicate output
The weight matrix that layer is connected with the output that high-level characteristic extracts part, i.e. the first weight matrix；" * " representing matrix multiplication, hfo table
Show that high-level characteristic extracts the output (output that each column indicate a sample) of part；w2Indicate output layer and low-level feature abstract
The connected weight matrix of partial output, i.e. the second weight matrix；Lfo indicates that the output matrix of low-level feature abstract part is (every
One column indicate the output of a sample)；B indicates output layer bias matrix.
Influence in view of two kinds of features to final result is added one in the reconstruction features vector for indicating low-level feature
Weight beta is adjusted, 0≤β≤1 is adjusted, and influence of the low-level feature to the judgement of final result obtains a degree of inhibition.
The then court verdict of final network output are as follows:
O=sigm (w1*hfo+β×w2*lfo+b) (4)
Step 2-1-4), got parms w using training method1、w2, b and β final value: artificial neural network, depth
The basic theories of study is divided into training two parts of parameter in the design and network of network.After designing the structure of network (i.e.
Obtain above-mentioned high-level characteristic, low-level feature and judgement formula), need training to determine that the value of each parameter in network (determines
w1、w2, the value and convolutional layer of b and β and the value of the parameter in down-sampling layer, each neuron in artificial neural network in L layers
Parameter value), then the face picture that identifies of the network handles could be used to carry out the applications such as actual classification and recurrence.
The training method takes error backpropagation algorithm, while in view of deep learning needs a large amount of training sample branch
It holds, to reduce calculated load, is trained in conjunction with stochastic gradient descent strategy.If being by all T training picture random divisions
Dry group, and every B one group (B wants that T can be divided exactly), total " T/B " group.It is above-mentioned designed will to own the input of " T/B " group in order
It in network, and then extracts high-level characteristic low-level feature and after carrying out gender judgement, completes primary training iteration；Then again by institute
Having T training picture random divisions is several groups, is still every B one group, altogether " T/B " group.It is random division every time, it is ensured that
Each group of picture is different from the last time after each division, and all " T/B " group is still inputted above-mentioned design in order
Network in, and then extract high-level characteristic low-level feature and after carrying out gender judgement, complete new primary trained iteration.In total into
Row E times trained iteration could finally obtain parameter w1、w2, b and β value.
Undated parameter w1、w2, detailed process is as follows by b:
Firstly, random initializtion parameter w1、w2, b and β value, then input first group of trained picture and obtain judgement to the end
As a result each column indicate that the judgement an of sample exports result in o, o.Then the error of output layer is calculated, calculation formula is such as
Under:
Wherein, MSE indicates square between court verdict o and actual sample label (classification, referred to herein as men and women) matrix Y
Error, MSE are matrix expression；Y is the label matrix of input sample, if the face picture of input is male, Y is just [1
0]TMatrix is then [0 1] if womenTMatrix；The court verdict that o exports for network, " | | | |2" corresponding between representing matrix
Matrix after the squared difference of element indicates.
Then, the parameter w in formula (4) can be calculated using this error1, w2, b and β are for final error letter
Number J (W, b；Gradient β) updates above-mentioned parameter w using gradient decline principle1, w2, the value of b and β.Wherein error function J (W,
b；Matrix form β) is MSE.For this purpose, the sensitivity of output layer need to be calculated:
Wherein, Od indicates the matrix form of output layer sensitivity,Dot product (corresponding element phase between representing matrix
Multiply, dimension is consistent).Wherein f ' (o) indicates that, to output function derivation, activation primitive uses sigmoid function, and derivative form is
F ' (x)=f (x) (1-f (x)).Parameter w can further be found out using following 3 formula according to sensitivity1, w2And biasing b
Value, i.e., to w1, w2And biasing b is updated:
Wherein, (w1)newIndicate the updated first weight matrix w in error back propagation each time1Value,
(w1)oldThe first weight matrix w before corresponding update1Value；α indicates the learning rate of network, and the present invention takes learning rate changing strategy
Training, i.e., the value of α is initialized as a biggish value, is then gradually reduced, guarantees whole with the increase of training the number of iterations
The convergence of a network.
Wherein, (w2)newIndicate the updated second weight matrix w in error back propagation each time2Value,
(w2)oldThe second weight matrix w before corresponding update2Value.
Wherein, (b)newIndicate the value of the updated bias matrix b in error back propagation each time, (b)oldIt is corresponding
The value of bias matrix b before update.
Detailed process is as follows by undated parameter β:
Since β is a real number, the more new formula for taking gradient descent method to update β is as follows:
Error function asks the formula of local derviation that can further be turned to by chain type rule β:
β and matrix (w2* lfo) matrix A and matrix (w of a ranks number identical as matrix can be regarded as by being multiplied2* lfo) point
Multiply, wherein element value is all β in A.In this way, formula (11) can finally turn to:
Wherein " mean (B (:)) " indicates to be averaged all elements in matrix B operation, and matrix B is indicated by each time
Error back propagation when the matrix that constitutes of the updated value for adjusting weight beta, the ranks number of the matrix B and court verdict o's
Ranks number is identical.
Error backpropagation algorithm, the network parameter w in achievable formula (4) are utilized in a word1, w2, the update of b and β.
It further include the parameter in convolutional layer in whole network, the parameter in parameter and L layer artificial neural network in down-sampling layer needs
Their value is determined by training.Institute's application method is still error backpropagation algorithm.Continue error MSE to forward pass
It broadcasts, joint voting layer a part and high-level characteristic extraction unit split-phase connect at this time, and a part is connected with network is rebuild, then MSE points
For two parts error, continue to propagate forward respectively in the two modules.The error back propagation of part is extracted in high-level characteristic
In the process, the side of the convolutional layer and down-sampling layer error back propagation undated parameter in existing depth learning technology is directlyed adopt
Method, ({ CS } indicates the matrix of all parameters in convolutional layer and down-sampling layer to the parameter { CS } for updating in convolutional layer and down-sampling layer
Set) value.During the error back propagation of low-level feature abstract part, existing artificial neural network technology is directlyed adopt
In error back propagation undated parameter method, update L layer artificial neural network in parameter { LN } ({ LN } expression L layers of people
The set of matrices of all parameters in artificial neural networks) value.It is then defeated this completes the parameter training process of one group of picture
When entering second group of picture training, with the determining network parameter w of first group of picture1, w2, the value of b and β, and the value of { CS } and { LN }
Court verdict o is calculated, then repetitive error back-propagation process, updates w in network1, w2, b and β value, and { CS } and { LN }
Value.And so on, the value of the parameter more than training of each group of picture determined after one group of picture training calculates court verdict, then
Error is calculated since output and propagates back to input, updates the value of all parameters of whole network.It is trained until " T/B " is organized
Sample fully enters network and completes after training, and just completes primary training iteration.
After completing all E trained iteration, network at this time can be used to actual gender identification, input face figure
Piece, network will export the predicted value of gender.
Two, the building of face age estimation network:
In above-mentioned technical proposal, the step 2) provides a kind of method of face age estimation, as shown in Fig. 2, i.e.
A kind of deep layer network system for the estimation of face age, comprising:
Step 2-2) using the training picture of face obtained in step 1) as the input of network, the output of network is in picture
The age estimated value (integer) of people.Network is mainly made of 4 partial function modules, including high-level characteristic extraction module, and low layer is special
Levy extraction module, fusion feature cascading judgement output module and parameter training module.
In above scheme, the step 2-1) include:
Step 2-2-1), high-level characteristic extract: directly adopt the volume of the convolutional neural networks in existing deep learning method
Lamination and down-sampling layer structure realize that the high-level characteristic for training picture extracts.In the specific implementation, using 3 layers of convolutional layer
C1, C2And C3And three layers of down-sampling layer S1, S2And S3Combination, full connection between layers.
Step 2-2-2), low-level feature abstract: it is synchronous with high-level characteristic extraction module that input picture is handled.
Firstly, the form for converting vector by two-dimensional graph structure form for the training picture of the face of input is (referred to as flat
Graduation processing, flat operation), then obtained vector is normalized, obtain face original feature vector.
Then, original feature vector is connected to and rebuilds network to obtain reconstruction features vector.Network is rebuild to be based on manually
Neural networks principles are built, L layers total.Each layer of neuron output is considered as feature vector, next layer of neuron pair
The vector of upper one layer of output is recompiled, and is indicated and is exported again again after describing, takes full connection between layers
Mode, each neuron activated using sigmoid function.Face original feature vector is first connected to reconstruction network
H1Layer, H1Layer contains h1A neural unit, it is assumed that input picture size is N × N, then face original feature vector is N × N
Dimension, by H1Become h after layer1Dimensional feature vector.It is then followed by and is connected to H2Layer, H2Layer contains h2A neural unit, then feature
Vector is further changed to h2Dimension.And so on, according to actual needs, it is finally coupled to HnLayer, obtains a hnThe feature of dimension to
Amount.The specific calculating of connection between layers can be obtained by formula (1).
This hnThe feature vector of dimension is to rebuild output namely the reconstruction features vector of network.Reconstruction features vector picks
In addition to estimating the age invalid redundancy in face original feature vector, reduce partial noise interference, it can be preferable
Portray low layer, the global characteristics of face.
Step 2-2-3), joint voting layer and output: the high-level characteristic and low-level feature of combined extracting carry out final year
Age estimation, output age predicted value (how old).
Obtained high-level characteristic figure will be extracted and be connected to the full articulamentum containing M neuron, obtain a M dimension
High-level characteristic vector.Then high-level characteristic vector, which links together to constitute with reconstruction features vector, combines voting layer.Assuming that rebuilding
Feature vector is hnDimension, then the two, which is united i.e. composition one, contains (M+hn) joint of a neuron decides by vote layer.Joint
The output of voting layer is the (M+h for the multiple features structure that our whole networks are extractedn) dimensional feature vector.
Joint voting layer is connected to final output neuron entirely again.Output neuron has S (each correspondence is one-year-old).
Output layer still uses sigmoid function to activate, then the probability of every one kind is represented by formula (2).Because whole network
Input plurality of pictures (parameter training part has respective description) every time, therefore the output matrix form of network that the present invention defines can
It is expressed as formula (3).
Influence in view of two kinds of features to final result is added one in the reconstruction features vector for indicating low-level feature
Weight beta is adjusted, 0≤β≤1 is adjusted, and influence of the low-level feature to the judgement of final result obtains a degree of inhibition.
Then final network output can be obtained by formula (4).
Step 2-2-4), got parms w using training method1、w2, b and β final value: artificial neural network, depth
The basic theories of study is divided into training two parts of parameter in the design and network of network.After designing the structure of network (i.e.
Obtain above-mentioned high-level characteristic, low-level feature and judgement formula), need training to determine that the value of each parameter in network (determines
w1、w2, the value and convolutional layer of b and β and the value of the parameter in down-sampling layer, each neuron in artificial neural network in L layers
Parameter value), then the face picture that identifies of the network handles could be used to carry out the applications such as actual classification and recurrence.
The training method takes error backpropagation algorithm, while in view of deep learning needs a large amount of training sample branch
It holds, to reduce calculated load, is trained in conjunction with stochastic gradient descent strategy.If being by all T training picture random divisions
Dry group, and every B one group (B wants that T can be divided exactly), total " T/B " group.It is above-mentioned designed will to own the input of " T/B " group in order
It in network, and then extracts high-level characteristic low-level feature and after carrying out gender judgement, completes primary training iteration；Then again by institute
Having T training picture random divisions is several groups, is still every B one group, altogether " T/B " group.It is random division every time, it is ensured that
Each group of picture is different from the last time after each division, and all " T/B " group is still inputted above-mentioned design in order
Network in, and then extract high-level characteristic low-level feature and after carrying out gender judgement, complete new primary trained iteration.In total into
Row E times trained iteration could finally obtain parameter w1、w2, b and β value.Undated parameter w1、w2, detailed process is as follows by b:
Firstly, random initializtion parameter w1、w2, b and β value, then input first group of trained picture and obtain judgement to the end
As a result each column indicate that the judgement an of sample exports result in o, o.Then the error of output layer is calculated using formula (5).
It should be noted that the sample label matrix Y in formula (5) is S dimensional vector form in age estimation, if input face
The picture corresponding age is 1 years old, then Y is then [1 0 ... 0]T；If the age is 2 years old, Y is [0 1 ... 0]T；If it is
S years old, then be [0 0 ... 1]T。
Then, the parameter w in formula (4) can be calculated using the error that formula (5) obtains1, w2, b and β are for most
Whole error function J (W, b；Gradient β) updates above-mentioned parameter w using gradient decline principle1, w2, the value of b and β.Wherein miss
Difference function J (W, b；Matrix form β) is MSE.For this purpose, the sensitivity of output layer need to be calculated, there can be formula (6) to obtain.Root
According to sensitivity, using formula (7), formula (8) and formula (9) can further find out parameter w1, w2And the value of biasing b, i.e., pair
w1, w2And biasing b is updated.
Since β is a real number, when gradient descent method being taken to update β, find out updated β's using formula (10)
Value.Error function asks the formula of local derviation that can further pass through formula (11) conversion by chain type rule β.Finally, β and matrix (w2*
Lfo the matrix A and matrix (w of a ranks number identical as matrix can be regarded as by) being multiplied2* lfo) dot product, wherein element value is all in A
For β.In this way, formula (11) can finally turn to formula (12), to calculate the updated value of β.
Error backpropagation algorithm, the network parameter w in achievable formula (4) are utilized in a word1, w2, the update of b and β.
It further include the parameter in convolutional layer in whole network, the parameter in parameter and L layer artificial neural network in down-sampling layer needs
Their value is determined by training, institute's application method is still error backpropagation algorithm.Continue error MSE to forward pass
It broadcasts, joint voting layer a part and high-level characteristic extraction unit split-phase connect at this time, and a part is connected with network is rebuild, then MSE points
For two parts error, continue to propagate forward respectively in the two modules.The error back propagation of part is extracted in high-level characteristic
In the process, the side of the convolutional layer and down-sampling layer error back propagation undated parameter in existing depth learning technology is directlyed adopt
Method, ({ CS } indicates the matrix of all parameters in convolutional layer and down-sampling layer to the parameter { CS } for updating in convolutional layer and down-sampling layer
Set) value.During the error back propagation of low-level feature abstract part, existing artificial neural network technology is directlyed adopt
In error back propagation undated parameter method, update L layer artificial neural network in parameter { LN } ({ LN } expression L layers of people
The set of matrices of all parameters in artificial neural networks) value.It is then defeated this completes the parameter training process of one group of picture
When entering second group of picture training, with the determining network parameter w of first group of picture1, w2, the value of b and β, and the value of { CS } and { LN }
Court verdict o is calculated, then repetitive error back-propagation process, updates w in network1, w2, b and β value, and { CS } and { LN }
Value.And so on, the value of the parameter more than training of each group of picture determined after one group of picture training calculates court verdict, then
Error is calculated since output and propagates back to input, updates the value of all parameters of whole network.It is trained until " T/B " is organized
Sample fully enters network and completes after training, and just completes primary training iteration.
After completing all E trained iteration, network at this time can be used to actual age estimation, input face figure
Piece, network will export the estimated value at age.
Above-mentioned formula (3), (4), (7), (8), (10), (12) are new formula proposed by the present invention.
Three, the estimation of face age and gender identification are carried out based on above-mentioned building network:
After constructing to obtain gender identification network and age estimation network using the above method, picture to be identified is inputted
Gender identification or age estimation, specific identification process are carried out in the network having been built up are as follows:
The high-level characteristic of face picture to be identified is extracted based on convolutional neural networks；
The low layer and global characteristics of face picture to be identified are extracted based on artificial neural network；
The low-level feature of extraction and high-level characteristic are inputted to following judgement formula, carry out gender or age judgement, it is defeated
Court verdict out:
O=sigm (w1*hfo+β×w2*lfo+b)
Wherein, the first weight matrix w in above-mentioned formula1, the second weight matrix w2, β and b be to complete for training picture
Iteration several times after obtain (the final value of each parameter obtained when i.e. above-mentioned network struction is completed), what hfo indicated to extract
The high-level characteristic vector of face to be identified, lfo indicate that the low-level feature vector of the face to be identified extracted, o indicate gender or year
The court verdict in age.
Low-level feature abstract further includes:
After the face picture to be identified of input is carried out flaky process and is normalized, obtain the primitive character of face to
Amount；
Original feature vector is inputted into artificial neural network, input vector is rebuild by multilayer neuronal structure,
Obtain a LfThe feature vector of dimension is as the low-level feature vector extracted.
Specific judging process are as follows:
The high-level characteristic and low-level feature extracted is merged, a kind of multiple features structure is exported；Using several output neurons
Gender judgement or age judgement are carried out, each neuron is based on the judgement formula and makes decisions.
The present invention gives a kind of system example of human face analysis method specific implementation based on multiple features deep learning,
As shown in Figure 3.Wherein " Face datection and the picture pretreatment " in Fig. 3 directlys adopt the realization of prior art means, Face datection
The method for combining the verifying of face regional area based on mark point detection scoring is realized；Picture pretreatment includes the gray scale of color image
Change processing, picture size adjustment and picture histogram equalization processing.System on human face picture is using such as after pretreatment
Lower strategy is identified:
High-level characteristic extraction module, for extracting the high-level characteristic of face picture to be identified based on convolutional neural networks；
Low-level feature abstract module, the low layer and the overall situation for extracting face picture to be identified based on artificial neural network are special
Sign；
Judging module, for the low-level feature of extraction and high-level characteristic to input to following judgement formula, carry out gender or
Person's age adjudicates, and exports court verdict:
O=sigm (w1*hfo+β×w2*lfo+b)
Wherein, the first weight matrix w in above-mentioned formula1, the second weight matrix w2, β and b be to complete for training picture
Iteration several times after obtain (the final value of each parameter obtained when i.e. above-mentioned network struction is completed), what hfo indicated to extract
The high-level characteristic vector of face to be identified, lfo indicate that the low-level feature vector of the face to be identified extracted, o indicate gender or year
The court verdict in age.
Above-mentioned low-level feature abstract module further includes:
Flaky process module, after the face picture to be identified for that will input carries out flaky process and normalizes,
Obtain the original feature vector of face；
Reconstruction features vector obtains module, for original feature vector to be inputted artificial neural network, passes through multilayer nerve
Meta structure rebuilds input vector, obtains a LfThe feature vector of dimension is as the low-level feature vector extracted.
Above-mentioned judging module further includes:
Joint voting layer module exports a kind of multiple features structure for merging the high-level characteristic and low-level feature that extract；
The judgement of final gender or age is carried out to output layer.
Output layer module, for carrying out gender judgement or age judgement, each mind using several output neurons
It is made decisions through member based on the judgement formula.
In short, the present invention provides a kind of human face analysis method and system based on multiple features deep learning.The system is first
Face datection first is carried out to picture and pretreatment obtains face picture, then face picture is analyzed.Human face analysis is specific
Estimate including gender classification and face age.System, which designs and trained the networks of two deep layers, is respectively used to face
Shi Bie not be with the estimation of face age, using face picture as the input of network, by the extraction of high-level characteristic, low-level feature is mentioned
Take and fusion feature after cascading judgement, two deep layer networks finally respectively export face picture gender prediction's value and the age
Estimated value.This system effectively raises the accuracy rate of Gender Classification and reduces the mean absolute error of age estimation, simultaneously
System has stronger generalization ability, can show on low-quality network picture and the picture of camera actual acquisition good
Performance.
The variable representing matrix of black matrix in above-mentioned technical proposal.
It should be noted last that the above examples are only used to illustrate the technical scheme of the present invention and are not limiting.Although ginseng
It is described the invention in detail according to embodiment, those skilled in the art should understand that, to technical side of the invention
Case is modified or replaced equivalently, and without departure from the spirit and scope of technical solution of the present invention, should all be covered in the present invention
Scope of the claims in.
Claims (9)
1. a kind of deep layer network establishing method estimated for gender classification or age, the method includes:
All training pictures are divided into several groups by step 101)；
Step 102) extracts the high-level characteristic of one group of trained picture based on convolutional neural networks, and then obtains by high-level characteristic vector
First matrix of composition；The low layer and global characteristics of same group of trained picture are extracted based on artificial neural network simultaneously, and then is obtained
The second matrix formed to low-level feature vector；
The first matrix, the second matrix and following judgement formula based on extraction obtain the knot of one group of gender identification or age estimation
Fruit:
O=sigm (w1*hfo+β×w2*lfo+b)
Wherein, hfo indicates the first matrix；Lfo indicates the second matrix；For the first power in first group of trained picture above-mentioned formula
Value matrix w1, the second weight matrix w2, bias matrix b and adjust the initial value of weight beta and obtained using random initializtion mode；For
W in the training picture above-mentioned formula of remaining each group of input1、w2, b and β value acquisition methods are as follows: reversely passed using error
Broadcast error function J (W, b that algorithm calculates the physical tags matrix Y of court verdict o and each group training picture；β), then pass through calculating
w1、w2, b and β be to error function J (W, b；Gradient and then undated parameter w β)1、w2, b and β value；
Step 103) inputs one group of trained picture again, and repeats above-mentioned steps 102 to the training picture inputted again), directly
Primary training iteration is completed in the processing for being carried out step 102) to all groupings；
All training pictures are reclassified as several groups by step 104), and repeat above-mentioned step to each group repartitioned
It is rapid 102) and step 103), complete iteration again；
By several grouping and iterative processings again, until being obtained final when the judgement o of final output meets the condition of setting
Parameter w1、w2, b and β value, complete network struction；
First weight matrix w is updated using the error backpropagation algorithm of following formula1Value:
Wherein, (w1)newIndicate the updated first weight matrix w in error back propagation each time1Value, (w1)oldIt is right
The first weight matrix w before should updating1Value, Od indicate output layer sensitivity matrix；The learning rate of α expression network；
The second weight matrix w is updated by following formula2Value:
Wherein, (w2)newIndicate the updated second weight matrix w in error back propagation each time2Value, (w2)oldIt is right
The second weight matrix w before should updating2Value.
2. the deep layer network establishing method according to claim 1 estimated for gender classification or age, feature
It is, further included when low-level feature abstract:
Step 102-11) by each Zhang Xunlian picture of one group of input trained picture from two-dimensional graph structure form be converted into
The form of amount, then vector is normalized, obtain the original feature vector of each Zhang Xunlian picture；
Step 102-12) by the original vector input artificial neural network of obtained each Zhang Xunlian picture, and then obtain one group
Reconstruction features vector to get arrive second matrix；Wherein, the artificial neural network includes L layers, and is adopted between layers
With full connection type, each layer of each neuron is activated using sigmoid function.
3. the deep layer network establishing method according to claim 1 estimated for gender classification or age, feature
It is, for the gender identification of a trained picture or age in the result for obtaining one group of gender identification or age estimation
Estimation procedure specifically includes:
Step 102-21) when extraction a trained picture high-level characteristic vector be HfThe high-level characteristic vector of dimension, and low layer is special
Sign vector is LfWhen the feature vector of dimension, construction one includes " Hf+Lf" joint of a neuron decides by vote layer；
Step 102-22) when for gender identification when, by construction joint voting layer each neuron respectively with output layer
Two output neurons are connected, and each output neuron is based on the judgement formula and carries out Sex Discrimination, exports training picture
For the probability of sex；When for age estimation, each neuron of joint voting layer and the S output mind of output layer
Through member be connected, wherein each output neuron correspond to it is one-year-old.
4. the deep layer network establishing method according to claim 1 estimated for gender classification or age, feature
It is, the value of the β is training update method in iteration each time are as follows:
Wherein, βnewIndicate the updated value for adjusting weight beta, β in error back propagation each timeoldBefore corresponding update
Adjust the value of weight beta；
The part of local derviation is asked in above-mentioned formula to be obtained by following formula:
Wherein, f ' (o) indicates that, to court verdict o derivation, " mean (B (:)) " indicates to be averaged fortune to all elements in matrix B
It calculates；Matrix B indicates the matrix being made of the value for adjusting weight beta updated in error back propagation each time, and the square
The ranks number of battle array B is identical as the ranks number of court verdict o.
5. a kind of for the age of face or the recognition methods of gender, this method is based on any one right in claim 1-4
It is required that the first weight matrix w that the building network recorded finally determines1, the second weight matrix w2, β and b value, the method packet
Contain:
The high-level characteristic of face picture to be identified is extracted based on convolutional neural networks；
The low layer and global characteristics of face picture to be identified are extracted based on artificial neural network；
The low-level feature of extraction and high-level characteristic are inputted to following judgement formula, gender is carried out or age judgement, output is sentenced
Certainly result:
O=sigm (w1*hfo+β×w2*lfo+b)
Wherein, the first weight matrix w in above-mentioned formula1, the second weight matrix w2, β and b be that deep layer network establishing method determines
Value, hfo indicate extract face to be identified high-level characteristic vector, lfo indicate extract face to be identified low-level feature
Vector, o indicate gender or the court verdict at age.
6. according to claim 5 for the age of face or the recognition methods of gender, which is characterized in that using following step
It is rapid to extract low-level feature:
After the face picture to be identified of input is carried out flaky process and normalized, the original feature vector of face is obtained；
Original feature vector is inputted into artificial neural network, input vector is rebuild by multilayer neuronal structure, is obtained
One LfThe feature vector of dimension is as the low-level feature vector extracted.
7. a kind of for the age of face or the identifying system of gender, which is characterized in that the system includes:
High-level characteristic extraction module, for extracting the high-level characteristic of face picture to be identified based on convolutional neural networks；
Low-level feature abstract module, for extracting the low layer and global characteristics of face picture to be identified based on artificial neural network；
Judging module neural network based, for the low-level feature of extraction and high-level characteristic to be inputted to following judgement formula,
Gender or age judgement are carried out, court verdict is exported:
O=sigm (w1*hfo+β×w2*lfo+b)
Wherein, the first weight matrix w in above-mentioned formula1, the second weight matrix w2If, β and b be to complete for training picture
It is obtained after dry iteration, hfo indicates that the high-level characteristic vector of the face to be identified extracted, lfo indicate the face to be identified extracted
Low-level feature vector, o indicates gender or the court verdict at age；
First weight matrix w is updated using the error backpropagation algorithm of following formula1Value:
Wherein, (w1)newIndicate the updated first weight matrix w in error back propagation each time1Value, (w1)oldIt is right
The first weight matrix w before should updating1Value, Od indicate output layer sensitivity matrix；The learning rate of α expression network；
The second weight matrix w is updated by following formula2Value:
Wherein, (w2)newIndicate the updated second weight matrix w in error back propagation each time2Value, (w2)oldIt is right
The second weight matrix w before should updating2Value.
8. the age of face according to claim 7 or the identifying system of gender, which is characterized in that the low-level feature mentions
Modulus block further includes:
Flaky process module obtains after the face picture to be identified for that will input carries out flaky process and normalizes
The original feature vector of face；
Reconstruction features vector obtains module and passes through multilayer neuron knot for original feature vector to be inputted artificial neural network
Structure rebuilds input vector, obtains a LfThe feature vector of dimension is as the low-level feature vector extracted.
9. the age of face according to claim 7 or the identifying system of gender, which is characterized in that the judging module into
One step includes:
Joint voting layer module exports a kind of multiple features structure for merging the high-level characteristic and low-level feature that extract；
Output layer module, for carrying out gender judgement or age judgement, each neuron using several output neurons
It is made decisions based on the judgement formula.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410193260.5A CN105095833B (en) | 2014-05-08 | 2014-05-08 | Network construction method, recognition method and system for face recognition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410193260.5A CN105095833B (en) | 2014-05-08 | 2014-05-08 | Network construction method, recognition method and system for face recognition |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105095833A CN105095833A (en) | 2015-11-25 |
| CN105095833B true CN105095833B (en) | 2019-03-15 |
Family
ID=54576227
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410193260.5A Active CN105095833B (en) | 2014-05-08 | 2014-05-08 | Network construction method, recognition method and system for face recognition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105095833B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107480582A (en) * | 2017-06-28 | 2017-12-15 | 北京五八信息技术有限公司 | The detection method and device of resume validity |
| US11966453B2 (en) | 2021-02-15 | 2024-04-23 | International Business Machines Corporation | Ordering annotation sets for machine learning |
Families Citing this family (44)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106803054B (en) | 2015-11-26 | 2019-04-23 | 腾讯科技（深圳）有限公司 | Faceform's matrix training method and device |
| CN105426963B (en) * | 2015-12-01 | 2017-12-26 | 北京天诚盛业科技有限公司 | For the training method of the convolutional neural networks of recognition of face, device and application |
| CN105678381B (en) * | 2016-01-08 | 2019-03-08 | 浙江宇视科技有限公司 | A gender classification network training method, gender classification method and related device |
| CN106991478B (en) * | 2016-01-20 | 2020-05-08 | 中科寒武纪科技股份有限公司 | Apparatus and method for performing artificial neural network reverse training |
| CN106991477B (en) * | 2016-01-20 | 2020-08-14 | 中科寒武纪科技股份有限公司 | Artificial neural network compression coding device and method |
| CN107103266B (en) * | 2016-02-23 | 2019-08-20 | 中国科学院声学研究所 | Two-dimensional face fraud detection classifier training and face fraud detection method |
| CN105844236B (en) * | 2016-03-22 | 2019-09-06 | 重庆医科大学 | Age testing method based on skin image information processing |
| CN109074499B (en) * | 2016-04-12 | 2020-06-02 | 北京市商汤科技开发有限公司 | Method and system for object re-identification |
| CN105975916B (en) * | 2016-04-28 | 2019-10-11 | 西安电子科技大学 | Age Estimation Method Based on Multi-Output Convolutional Neural Network and Ordinal Regression |
| CN106127159A (en) * | 2016-06-28 | 2016-11-16 | 电子科技大学 | A kind of gender identification method based on convolutional neural networks |
| CN107657204A (en) * | 2016-07-25 | 2018-02-02 | 中国科学院声学研究所 | The construction method and facial expression recognizing method and system of deep layer network model |
| CN106295521B (en) * | 2016-07-29 | 2019-06-04 | 厦门美图之家科技有限公司 | A kind of gender identification method based on multi output convolutional neural networks, device and calculate equipment |
| CN107688576B (en) * | 2016-08-04 | 2020-06-16 | 中国科学院声学研究所 | A CNN-SVM model construction and propensity classification method |
| CN107766774A (en) * | 2016-08-17 | 2018-03-06 | 鸿富锦精密电子(天津)有限公司 | Face identification system and method |
| CN107871106B (en) * | 2016-09-26 | 2021-07-06 | 北京眼神科技有限公司 | Face detection method and device |
| CN106485235B (en) * | 2016-10-24 | 2019-05-03 | 厦门美图之家科技有限公司 | A kind of convolutional neural networks generation method, age recognition methods and relevant apparatus |
| CN106503661B (en) * | 2016-10-25 | 2019-06-28 | 陕西师范大学 | Face gender identification method based on fireworks deepness belief network |
| CN108229263B (en) * | 2016-12-22 | 2021-03-02 | 杭州光启人工智能研究院 | Target object identification method and device and robot |
| CN106874929B (en) * | 2016-12-28 | 2020-03-17 | 浙江诸暨会智科技有限公司 | Pearl classification method based on deep learning |
| CN106991147A (en) * | 2017-03-27 | 2017-07-28 | 重庆大学 | A kind of Plant identification and recognition methods |
| CN108734719A (en) * | 2017-04-14 | 2018-11-02 | 浙江工商大学 | Background automatic division method before a kind of lepidopterous insects image based on full convolutional neural networks |
| CN107239532B (en) * | 2017-05-31 | 2020-07-31 | 北京京东尚科信息技术有限公司 | Data mining method and device |
| CN107545238A (en) * | 2017-07-03 | 2018-01-05 | 西安邮电大学 | Underground coal mine pedestrian detection method based on deep learning |
| CN107463965B (en) * | 2017-08-16 | 2024-03-26 | 湖州易有科技有限公司 | Deep learning-based fabric attribute picture acquisition and recognition method and recognition system |
| CN107506737A (en) * | 2017-08-29 | 2017-12-22 | 四川长虹电器股份有限公司 | Face gender identification method |
| CN107545249A (en) * | 2017-08-30 | 2018-01-05 | 国信优易数据有限公司 | A kind of population ages' recognition methods and device |
| CN107622282A (en) * | 2017-09-21 | 2018-01-23 | 百度在线网络技术（北京）有限公司 | Image verification method and apparatus |
| CN107844760A (en) * | 2017-10-24 | 2018-03-27 | 西安交通大学 | Three-dimensional face identification method based on curved surface normal direction component map Neural Networks Representation |
| CN108052862B (en) * | 2017-11-09 | 2019-12-06 | 北京达佳互联信息技术有限公司 | Age estimation method and device |
| CN108171325B (en) * | 2017-12-28 | 2020-08-28 | 中国科学技术大学 | A time series integrated network, encoding device and decoding device for multi-scale face restoration |
| CN108182429B (en) * | 2018-02-01 | 2022-01-28 | 重庆邮电大学 | Method and device for extracting facial image features based on symmetry |
| CN110502959A (en) * | 2018-05-17 | 2019-11-26 | Oppo广东移动通信有限公司 | Gender judgment method, gender judgment device, storage medium and electronic equipment |
| CN109002755B (en) * | 2018-06-04 | 2020-09-01 | 西北大学 | Age estimation model construction method and estimation method based on face image |
| CN109001833A (en) * | 2018-06-22 | 2018-12-14 | 天和防务技术（北京）有限公司 | A kind of Terahertz hazardous material detection method based on deep learning |
| CN108833713A (en) * | 2018-06-28 | 2018-11-16 | Oppo广东移动通信有限公司 | Antenna control method during sleep and related product |
| CN109344840B (en) * | 2018-08-07 | 2022-04-01 | 深圳市商汤科技有限公司 | Image processing method and apparatus, electronic device, storage medium, and program product |
| CN110956190A (en) * | 2018-09-27 | 2020-04-03 | 深圳云天励飞技术有限公司 | Image recognition method and device, computer device, and computer-readable storage medium |
| CN109711243B (en) * | 2018-11-01 | 2021-02-09 | 长沙小钴科技有限公司 | A Deep Learning-Based Static 3D Face Liveness Detection Method |
| WO2020093306A1 (en) * | 2018-11-08 | 2020-05-14 | 北京比特大陆科技有限公司 | Neutral network layer grouping method, apparatus and device, and storage medium and program product |
| CN109872364B (en) * | 2019-01-28 | 2022-02-01 | 腾讯科技（深圳）有限公司 | Image area positioning method, device, storage medium and medical image processing equipment |
| CN111723613A (en) * | 2019-03-20 | 2020-09-29 | 广州慧睿思通信息科技有限公司 | Face image data processing method, device, equipment and storage medium |
| CN110378372A (en) * | 2019-06-11 | 2019-10-25 | 中国科学院自动化研究所南京人工智能芯片创新研究院 | Diagram data recognition methods, device, computer equipment and storage medium |
| CN111753641B (en) * | 2020-05-07 | 2023-07-18 | 中山大学 | A Gender Prediction Method Based on High Dimensional Features of Face |
| CN113298156B (en) * | 2021-05-28 | 2024-09-06 | 有米科技股份有限公司 | Neural network training method and device for classifying image sexes |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102612841A (en) * | 2009-11-17 | 2012-07-25 | Lg电子株式会社 | Method for user authentication, and video communication apparatus and display apparatus thereof |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5359699A (en) * | 1991-12-02 | 1994-10-25 | General Electric Company | Method for using a feed forward neural network to perform classification with highly biased data |
| CN102339384B (en) * | 2011-09-16 | 2013-07-03 | 北京交通大学 | Face recognition method |
| CN103679185B (en) * | 2012-08-31 | 2017-06-16 | 富士通株式会社 | Convolutional neural networks classifier system, its training method, sorting technique and purposes |
| CN103778414A (en) * | 2014-01-17 | 2014-05-07 | 杭州电子科技大学 | Real-time face recognition method based on deep neural network |
-
2014
- 2014-05-08 CN CN201410193260.5A patent/CN105095833B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102612841A (en) * | 2009-11-17 | 2012-07-25 | Lg电子株式会社 | Method for user authentication, and video communication apparatus and display apparatus thereof |
Non-Patent Citations (2)
| Title |
|---|
| Robust Pedestrian Detection Using Low Level and High Level Features;Fariba Takarli，Ali Aghagolzadeh，Hadi Seyedarabi;《2013 21st Iranian Conference on Electrical Engineering》;20130916;第2,3,4,6章 * |
| 典型人工神经网络;巩敦卫，孙晓燕;《智能控制技术简明教程》;20101231;第127-128页 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107480582A (en) * | 2017-06-28 | 2017-12-15 | 北京五八信息技术有限公司 | The detection method and device of resume validity |
| US11966453B2 (en) | 2021-02-15 | 2024-04-23 | International Business Machines Corporation | Ordering annotation sets for machine learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105095833A (en) | 2015-11-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105095833B (en) | Network construction method, recognition method and system for face recognition | |
| CN113011499B (en) | Hyperspectral remote sensing image classification method based on double-attention machine system | |
| Zhao et al. | A deep learning iris recognition method based on capsule network architecture | |
| Suryawati et al. | Deep structured convolutional neural network for tomato diseases detection | |
| CN112801040B (en) | Lightweight unconstrained facial expression recognition method and system embedded with high-order information | |
| CN107122809A (en) | Neural network characteristics learning method based on image own coding | |
| Murugan | Feed forward and backward run in deep convolution neural network | |
| CN107423756A (en) | Nuclear magnetic resonance image sequence sorting technique based on depth convolutional neural networks combination shot and long term memory models | |
| CN115644823A (en) | Dynamic prediction and individualized intervention method and system for rehabilitation effect | |
| CN107657204A (en) | The construction method and facial expression recognizing method and system of deep layer network model | |
| CN105160400A (en) | L21 norm based method for improving convolutional neural network generalization capability | |
| CN116665300B (en) | Skeleton action recognition method based on space-time self-adaptive feature fusion graph convolution network | |
| CN112818764A (en) | Low-resolution image facial expression recognition method based on feature reconstruction model | |
| Wang et al. | MLFC-net: A multi-level feature combination attention model for remote sensing scene classification | |
| CN115439754A (en) | A large-scale crop mapping method across phenological regions based on time-series remote sensing images | |
| CN115601583A (en) | A Deep Convolutional Network Object Recognition Method with Dual-Channel Attention Mechanism | |
| CN109002755A (en) | Age estimation model building method and estimation method based on facial image | |
| CN105404865A (en) | Probability state restricted Boltzmann machine cascade based face detection method | |
| Lin et al. | Lateral refinement network for contour detection | |
| CN113706404B (en) | A method and system for correcting face images at depressed angles based on self-attention mechanism | |
| CN114170657A (en) | Facial emotion recognition method integrating attention mechanism and high-order feature representation | |
| CN114842542A (en) | Facial action unit identification method and device based on self-adaptive attention and space-time correlation | |
| Da et al. | Brain CT image classification with deep neural networks | |
| CN118537727A (en) | Hyperspectral image classification method based on multi-scale cavity convolution and attention mechanism | |
| Zhao et al. | Capsule networks with non-iterative cluster routing |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20211119 Address after: Building 5, innovation park, Keji East Road, wulongjiang middle Avenue, Shangjie Town, Minhou County, Fuzhou, Fujian 350108 Patentee after: HENGFENG INFORMATION TECHNOLOGY CO.,LTD. Address before: 100190, No. 21 West Fourth Ring Road, Beijing, Haidian District Patentee before: INSTITUTE OF ACOUSTICS, CHINESE ACADEMY OF SCIENCES |
|
| TR01 | Transfer of patent right |